AI gave junior auditors better judgment and not just louder voices

Andrew Jones RN — Assistant Professor of Accounting Andrew Jones says that generative AI helps junior auditors distinguish real concerns from false alarms.

Have you ever typed a question into ChatGPT that you’d never ask a colleague, because you felt you should already know the answer? Most of us likely have. Generative AI gives us something no co-worker can: information tailored to our exact situation, available before we have to commit to anything in front of anyone. For businesses built on expertise, that raises two questions. Does this kind of on-demand, tailored information actually improve how professionals act? And when it does (or doesn’t), what is it actually doing — making people bolder, or making them better?

My co-authors and I studied this with professional auditors. An audit typically depends on junior staff members noticing when something looks “off” and then telling their supervisor. They often don’t, because raising a nonissue carries real reputational costs, and junior auditors frequently can’t tell whether they’ve uncovered an issue or a nonissue. Furthermore, junior auditors can’t test the waters with their supervisors, because asking a supervisor is speaking up. When working in person, they can manage this risk by reading social cues and softening how the issue is raised. But when working remotely, they lose those social cues, and prior research shows that speaking up declines.

In our experiment, we gave some junior auditors a ChatGPT-style explanation of what their evidence implied and gave other junior auditors nothing, and we varied whether they were working in person or remotely. Without AI, the negative effect of remote work appeared as expected. With AI, it disappeared: Remote auditors with AI spoke up at similar levels as those in person. The telling result came from a second scenario with no real issue in it. There, generative AI information made auditors less likely to speak up. The AI wasn’t emboldening them; it was helping them tell real issues from false alarms, doing the job that in-person cues had been doing prior to generative AI.

Moving forward, our research suggests that the value of generative AI for this type of work may lie less in automating tasks than in supplying judgment-relevant understanding at the moment before the decision, when human experts are either unavailable or too costly to consult. But the value exists only if professionals choose to use the tool, prompt it well (e.g., include all the necessary details), and get accurate answers back. On that first condition, our next study asks whether auditors actually consult generative AI when it might tell them something they’d rather not hear. I suspect the answer will be humbling.

Bhaskar, L.S., A.K. Jones, and K. Kadous. 2026. “An Investigation Into How and When Generative AI Can Improve Auditors’ Voice Decisions.” Working paper, The University of Texas at Austin.

Back to November 2026 Research Newsletter